Domain Anomaly Detection in Machine Perception: A System Architecture and Taxonomy

Area of Science:

• Machine Perception
• Artificial Intelligence
• Computer Vision

Background:

• Traditional anomaly detection methods often fail to capture the nuances of domain-specific anomalies.
• A clear taxonomy and unified framework are needed to address the multifaceted nature of anomalies in machine perception.

Purpose of the Study:

• To introduce a novel concept of domain anomaly, distinct from conventional anomalies.
• To propose a unified Bayesian probabilistic framework for identifying and distinguishing various facets of domain anomalies.
• To develop a taxonomy of domain anomaly events.

Main Methods:

• The framework utilizes Bayesian probabilistic reasoning to define and differentiate concepts like outliers, noise, distribution drift, novelty detection, rare events, and unexpected events.
• A key mechanism involves detecting incongruence between contextual and non-contextual sensor data interpretation.
• The methodology is applied to anomaly detection in video annotation systems.

Main Results:

• The proposed framework successfully categorizes and identifies different types of domain anomalies.
• The unified approach underpins various existing anomaly detection applications.
• Demonstrated effectiveness in anomaly detection for video annotation.

Conclusions:

• The developed framework offers a comprehensive approach to domain anomaly detection in machine perception.
• The Bayesian probabilistic reasoning provides a robust foundation for understanding and classifying anomalies.
• This methodology has broad applicability across diverse machine perception tasks.